代理基礎設施
NemoClaw 0.0.93 Improves Local Agent Installation on DGX, Makes Download Credentials Ephemeral, and Adds Resume Support
NVIDIA has updated NemoClaw’s installation process for DGX Station and DGX Spark to prevent existing vLLM workloads from being stopped accidentally and to improve credential and rate-limit handling for large model downloads. The release remains in alpha, and the newly supported DGX OS configurations have not yet completed end-to-end qualification.
NVIDIA released NemoClaw 0.0.93 on July 23. Rather than adding new agent capabilities, the release focuses on fixing the most failure-prone installation and recovery paths in the local agent stack. NemoClaw is an Apache 2.0-licensed reference stack that runs OpenClaw, Hermes, or LangChain Deep Agents Code inside an OpenShell sandbox and uses a single CLI to manage inference providers, network policies, and lifecycle operations.
The new release enables DGX Station Express to recognize factory-installed, non-OTA DGX OS 7.6.x systems in the GB300 series while retaining checks for hardware, drivers, ECC, Docker, and CDI. When the platform generation cannot be verified, it continues to fail closed. If the installer detects an existing vLLM process, it no longer stops the workload automatically by default. Instead, it prints the exact command for stopping it manually and a version-pinned command for resuming the installation. Non-interactive mode likewise preserves the current state, reducing the risk that automated deployments will overwrite existing services.
For large model downloads, NemoClaw now explains the optional Hugging Face read token, passes the token only to the temporary download process, and sanitizes standard output and standard error. When an HTTP 429 response occurs, the recovery procedure reuses the existing cache instead of downloading all model weights again. On DGX Spark, when no provider is specified, the new priority order is an existing local vLLM instance, managed vLLM, and finally NVIDIA Endpoints. This changes the default inference location and data path for unattended installations.
These changes warrant attention from teams running self-hosted agents because the installer itself controls download credentials, containers, and GPUs, while its failure semantics also form part of the security boundary. However, NVIDIA explicitly labels the project as alpha, and full Station Express validation for DGX OS 7.6.x is still underway. Before upgrading, teams should preserve sandbox state; verify exit codes, agent network policies, and local model selection; and confirm that the new automatic selection order does not bypass their organization’s existing data-governance settings.